Note Steward

engrams: Cortex open-sources the always-on cloud factory around Claude Code and Codex

Sep 24 Unni: open-source engrams runs agents in your cloud — scheduled CVE/dep PRs, Datadog hotspot rewrites, error triage with dedupe, Linear+Slack end-to-end with microVM sub-agents.

Source: Introducing engrams: an open source software factory

Nikhil Unni’s Sep 24 Cortex builders note open-sources engrams — a platform for running Claude Code, Codex, or your own harness in your cloud (Firecracker / nested virt on GKE or EKS; Apple Virtualization or KVM locally). Internally, Cortex already uses it daily via UI or Slack mention, but the claim is that scheduled automations are where cloud agents beat interactive peak sessions. You bring Anthropic/OpenAI keys or route through OpenRouter so each workload picks a cost-appropriate model.

Four factory loops they describe as live: (1) Dependency / CVE upgrades on a schedule — open each upgrade, run tests, fix breakages, one PR per change, with a review automation before a human; escalate when a runtime jump outruns the suite. (2) Memory hot spots — daily Datadog profile query → task on the owning service → rewrite with before/after numbers in the PR. (3) Bug triage — every new runtime error gets reproduce/diagnose, then a fix or a rooted issue; repeat fires join the in-flight run instead of spawning duplicates; noise is named as noise rather than forced into a PR. (4) End-to-end delivery experiment — one long-running agent owns a Linear project and a Slack channel, picks issues, opens PRs, answers threads, posts a daily status, and babysits failing CI until green, spawning microVM sub-agents per ticket. Repo: cortexapps/engrams (0.x; APIs will move).

Why it matters for code review

Always-on factories that only mint more diffs overnight are optimizing the wrong layer — engrams’ useful unit is a verified automation with an escalation path, not a longer Claude Code burst. Review continuous cloud runs the way you review a merge queue: one PR per upgrade, before/after receipts on perf PRs, deduped error work, and explicit human gates when the suite cannot speak. Steady harnessed throughput compounds when the night owns the janitor and triage loops; FOMO for idle agents is wasted if daytime heroes still do the CVE grind by hand.